AI systems can generate code quickly but may introduce vulnerabilities or rely on outdated practices, highlighting that all AI systems are fundamentally code and can be exploited.
AI systems are prone to classic failures like overfitting and can be maliciously manipulated through deepfakes, chatbot poisoning, and adversarial patterns.
Several organizations like OWASP, NIST, and MITRE provide threat models and standards to help developers understand and mitigate AI security risks.
Attack trees categorize novel threats like evasion with adversarial samples, data poisoning to create backdoors, and model stealing to replicate proprietary systems.
A live demonstration shows how pre-generated adversarial samples can trick a digit recognition model into misclassifying numbers as zero.
Security risks extend beyond the model to the supply chain, including backdoors in pre-trained models, insecure serialization formats like Pickle, and vulnerabilities in ML frameworks.
The HDF5 format is recommended as a safer, industry-standard alternative to Python's insecure Pickle format for serializing machine learning models.
Wilken GmbH
Ulm, Germany
27:19Nura Kawa
27:02Mirko Ross
31:19Wolfgang Ettlinger & Alexander Hurbean
35:16Ankit Patel
27:26Cassie Kozyrkov
30:36Mackenzie Jackson
30:02Martin Förtsch & Thomas Endres
27:32Sebastian Schrittwieser
Jobs that call for the skills explored in this talk.
Wolters Kluwer
Alphen aan den Rijn, Netherlands
Collaboration Betters The World GmbH
Apple's Security Engineering & Architecture
Hugging Face
Paris, France
Ai-gruppe
München, Germany
